Fix HIP graph launch and runtime edge semantics

[AWoloszyn]

Allow no-argument native HIP kernel launches to be represented consistently across graph and direct stream paths when host stubs provide an empty args-array. Non-empty args-array launches without parameter metadata still fail explicitly instead of inventing an ABI layout.

Tighten graph memcpy and runtime pointer semantics around pooled logical allocation sizes, zero-byte allocation behavior, host-pointer attributes, worker-thread stream copies, event barriers, host-visible allocation backing, and native argument ABI offsets.

[zjgarvey]

Looking at some other callers for iree_hip_validate_memcpy_ranges, I noticed that the hipMemcpy and hipMemcpyAsync pass false for use_hip_visible_pool_size, and wasn't sure if those direct stream paths should use this logical size bound, too.

[AWoloszyn]

Thanks, these were valid cleanups. I added follow-up commit 6d437bceb.

• hipMemcpy and hipMemcpyAsync now use the HIP-visible logical allocation bound, matching the graph memcpy path. That matters for pooled allocations where the backing allocation can be larger than the
  allocation visible through HIP.
• The duplicated parameter-info empty predicate is now shared. I used it in the pack/unpack paths, but kept the separate “requires parameter storage” check because direct-argument-only metadata is a valid
  state and should not require a packed/list parameter buffer.
• src_ordinal was redundant; the metadata loop does not skip parameters, so it now uses j directly.